Robotlar
Industry Guide

Teaching a Go2 to Walk from Scratch: A Review of the go2-rl Project

UNITREE Turkey Hub & Competence Center
Published under the MIT license by an independent developer, the go2-rl project shows end-to-end how a Unitree Go2 learns walking, running, lying down/getting up and LiDAR obstacle avoidance from random initialization via reinforcement learning. As the Turkey distributor we reviewed it: what does it mean for teams evaluating the Go2 as a research platform, and how do you try it on your own robot?

What Is the go2-rl Project?

go2-rl is an educational open-source project published under the MIT license by independent developer Macmachi. It trains the 12-joint base configuration of the Unitree Go2 in the Genesis 1.2 simulator using PPO from the rsl-rl library: across 4,096 parallel simulation environments, the robot discovers walking on its own from zero knowledge. Training progress is documented with recordings showing behavior videos side by side with learning curves.

Why It Stands Out: No NVIDIA Required

Most RL training stacks depend on CUDA — including the official unitree_rl_gym (Isaac Gym-based, our guide here). go2-rl instead runs entirely on AMD Radeon + ROCm. For labs with tight GPU budgets or NVIDIA supply issues, that opens an alternative door into quadruped RL. Read the two tools as complementary, not competing: go2-rl for prototyping and teaching curricula, unitree_rl_gym for the official sim-to-real pipeline.

Sim-to-Real Discipline: Real Motor Limits

The project's standout trait is enforcing real Go2 hardware limits in simulation: joint torque limits (23.7 / 35.55 N·m) and velocity limits (30.1 / 20.07 rad/s) enter training as hard constraints, with domain randomization on top. That directly targets the classic 'flies in sim, falls on the real robot' failure mode — compare with the hardware realities we cover in our Unitree joint motors guide.

Which Go2 to Try It On: The EDU Difference

You don't need a robot to train in simulation; transferring the policy to a real robot requires a model with developer access. Go2 EDU is the right variant, with SDK access for secondary development and extended sensor options; the standard consumer Go2 does not expose low-level motor control. Setting up a university lab? See our procurement guide too.

Go2 Supply and Support in Turkey

As the authorized Turkey distributor we supply all variants including Go2, Go2 EDU and Go2-W with official warranty, installation and training support. We help you pin down the right configuration (sensor package, SDK access, battery options) for your research project and quote accordingly; the current Go2 pricing guide is here.

Robotlar.org Note

go2-rl is an independent developer's project; neither Robotlar.org nor Unitree Robotics is affiliated with it. This page is a distributor's review of the project's public documentation; use the project's GitHub repository for technical questions and contact us for Go2 supply. We track other open-source Unitree tools on our open source page.

Let's Find Your Solution

Our expert engineering team is ready to evaluate your project.